Sometimes you have to get creative – not just when writing lyrics, but when outsmarting algorithms, too. My new track "Acht-Drei" (German for "Eight-Three") is an emotional trip back to the year I was born, 1983. It's about the era of cassette tapes you'd wind back into shape with a pencil, about growing up without smartphones, and about the feeling of being the bridge between the analogue past and the digital present.
What makes this project special: it's not just the video that's AI-generated, the whole piece is a co-creation with artificial intelligence. The basic idea and the emotional cornerstones of the lyrics are entirely mine, but Google Gemini took care of the fine-tuning and wrote the final lines. And that's not all – Gemini then also spat out the highly specialised prompt optimisation for the music AI suno.com, which in the end generated the complete audio track.
For that honest vibe I wanted to pull off a very special music video: a real, uncut one-take video. The camera was supposed to stay in the recording studio and capture nothing but the performance in front of the microphone. The starting point for this experiment was a simple profile photo of me.
The mammoth project is being built with the video AI LTX 2.3. The plan was simple: take my real face and use image-to-video to put it into a professional studio setting. That first studio start frame was also edited with Gemini to put me in the right light. And in that first still, it worked surprisingly well.
But anyone who has worked with image-to-video models knows how frustrating it can get: visual continuity for photorealistic faces over longer stretches is an absolute disaster. If the rapper moves too wildly or turns his head away even slightly, the AI forgets who he is. As soon as he turns back, the model suddenly invents a completely new face.
After countless failed attempts, I threw the concept out entirely: we're leaving live action behind and switching to a three-dimensional comic and graphic-novel style. The genius part? The AI understands the clean geometry, edges and shading of a 3D comic character much better and stays visually stable across the whole video. On top of that, this look lets me move much more dynamically in front of the mic, like a rapper, without the AI distorting my face.
Using the last-frame method, the video is split into 19 segments in total (10 seconds each, with a 2-second finale). The last still of each scene is always the starting point for the next. Since I'm using a local software alternative on my own machine for the actual editing and rendering, I have to be extremely precise when writing the prompts for LTX 2.3 so I don't waste precious compute time. Here, too, I turn to Gemini: in the chat I feed the AI the last frame along with the exact lyric lines for the next scene. Gemini turns that into highly optimised English LTX prompts that control exactly which words get lip-synced and which LoRA strengths (for static control, detailer or Talkvid) need to be active.
But the very first scene already had a huge problem: for the first six seconds there are only ad-libs and the beat before the actual lyrics kick in. In stretches like that, the lip-sync AI desperately tries to turn every background noise into mouth movement. It looked extremely unnatural. My solution? A visual trick that I once again edited into the frame with Gemini: I put a glowing joint in my animated self's hand.
In the local LTX prompt I enforced a strict timeline: for the first six seconds the mouth stays completely shut as far as lip sync goes. Instead, the character raises the joint to his lips, takes a deep drag while nodding casually to the beat, and blows out the smoke. Only in the last four seconds does he switch into the rap performance and line up the precise lip sync for the first real words, "Acht-Drei". So a technical problem turned into a damn cool stylistic element that fits the vibe of the track perfectly.
All in all, I'm expecting around 20 hours of pure generation time on my local machine, during which countless clips will end up in the bin because of small glitches or new ideas. It's constant iterating and refining. But that's exactly what real commitment and passion for the art look like – then as now. Stay tuned for the finished video!